Short and Medium-Term Campus Electricity Load Forecasting with the CNN-Transformer-LSTM Hybrid Model

Accurate power load forecasting is critical for effective energy management in a campus environment where power use is affected by dynamic factors such as campus schedules, events, and weather. In order to improve the accuracy of short-to-medium term campus load prediction, this paper proposes a hybrid model, which combines Convolutional Neural Network (CNN) for local feature extraction and Transformer encoder for capturing global dependencies. Long Short-Term Memory (LSTM) Network is utilized to model long-term time patterns. The experimental results show that the CNN-Transformer-LSTM model consistently outperforms both the CNN-LSTM and conventional LSTM models, achieving significantly lower RMSE (57.5259), MAE (44.7930), and MAPE (1.05%) values, especially in medium-term prediction, with significant improvement in short-term prediction as well (RMSE: 44.6703, MAE: 36.5093, MAPE: 0.96%). This hybrid model demonstrated excellent learning capabilities in capturing complex time dependencies, making it an effective and powerful solution for enhancing campus energy management and strategic planning.